<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0409">
  <Title>Exceptionality and Natural Language Learning</Title>
  <Section position="5" start_page="5" end_page="5" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper we attempted to generalize the results of a previous study to a new set of language learning tasks from the area of spoken dialog systems. Our experiments indicate that previous results do not generalize so obviously to the new tasks. Next, we showed that some exceptionality measures can be used as means to improve the prediction accuracy on our tasks by combining the prediction of our learners based on measures of instance exceptionality. We observed that our memory-based learner performs better than the rule-based learner on typical instances and they exchange places for exceptional instances. We also showed that there is potential for moving these results from offline to online by performing a simple interpolation. Future work needs to address more complicated methods of interpolation, comparison between our method and other attempts to combine rule-based learning and memory-based learning (Domingos, 1996; Golding and Rosenbloom, 1991), comparison with ensemble methods, and whether the results from this paper generalize to other spoken dialog corpora.</Paragraph>
  </Section>
class="xml-element"></Paper>